Oracle Corp. is expanding the scope of its AI Agent Studio for Fusion Applications platform for building, testing and deploying artificial intelligence agents in one of a series of announcements at a sold-out Oracle AI World this week.

The company said today it’s doubling down on its belief that the enterprise future of AI is about operationalizing models, and it’s introducing features that enhance governance, security and management visibility into agents during construction and deployment.

“We think that the year 2026 will be the year companies start to really operationalize AI within their organizations,” said Natalia Rachelson, Oracle’s senior vice president of cloud applications development.

Launched last March, AI Agent Studio provides a no-code, Fusion-native environment for building, testing, managing and evaluating AI agents and agentic workflows. The updated version features new deep monitoring and evaluation capabilities, a real-time dashboard, token usage tracking, telemetry for agent behavior and built-in evaluation tools.

Agents in control

Customers can track how often agents succeed or fail, what prompts have been modified, which credentials have been added and where humans were injected into workflows. The intention is to bring AI agents in line with existing governance standards that organizations use for software-as-a-service applications.

Agents are “no different than other enterprise SaaS software,” Rachelson said. “That’s how we’re treating the studio and the agents within the studio,” she said.

Among the most notable upgrades is support for deterministic agent workflows, a feature that blends AI flexibility with operational predictability. They allow users to define a sequence of actions that an agent will execute exactly as instructed.

The feature is intended to address one of the most often-cited concerns about AI models, which is that their probabilistic reasoning makes them behave unpredictably. The enhanced Studio allows developers to create step-by-step workflows.

“The agent executes the steps and reasons as it’s executing the steps,” Rachelson said. “It’s very predefined and very controlled. It will deliver the same results 100% of the time.”

Deterministic agents can include loops, branching logic and human-in-the-loop approvals integrated into Fusion workflows.

“Oracle is making AI easier to adopt and harder to avoid,” Rebecca Wettemann, CEO of industry analyst firm Valoir, told SiliconANGLE in an email. “It’s important to note that unlike many of its competitors, Oracle is not charging for AI agents or for use of Agent Studio. This means customers have more flexibility to both experiment with a broadly deployed AI in production without worrying about budgetary constraints.”

LLM flexibility

Oracle is also expanding support for large language models to include Anthropic PBC’s Claude, Cohere Inc.’s Command, Google LLC’s Gemini, Meta Platforms Inc.’s Meta, OpenAI’s ChatGPT and x.AI Corp.’s Grok. Users can mix and match models within the same use case.

“When dealing with text and multimodal videos and images, you may prefer to use a different model,” Rachelson said. “Customers they can pick any LLM for any use case.”

The studio also introduces cost monitoring across models, including tracking of token usage for premium models like OpenAI’s GPT-4, which doesn’t currently run on Oracle Cloud Infrastructure and thus incurs extra cost.

“We provide a very generous allotment of tokens that the customers can start with,” Rachelson said. “Once they pass the threshold, that’s where the counting begins.”

A consolidated dashboard lets customers monitor and control spending “such that it’s as predictable as possible,” she said. “If they want to turn off the spigot, they can.”

Oracle said new prompt and topic libraries within AI Agent Studio ensure that prompts are consistent, reusable and governed. “Agents are tied to Fusion, so they adhere to Fusion’s security framework and Fusion access controls,” Rachelson said. “We know everything about the user: what their request center is, what legal entity they belong to, who their manager is and what their job code is. They can’t just access sensitive finance or payroll information.”

Marketplace of agents

Oracle will continue to release new prebuilt agents across a range of enterprise functions. This week, it’s introducing an AI Agent Marketplace and access to a growing network of more than 32,000 experts certified in agent development and deployment.

Certification requirements includes Oracle University training and a review of submitted work. “We check everything,” Rachelson said.

About 5,000 customers are testing or deploying agents using AI Agent Studio. Though Rachelson said it’s too early to talk about the results they’re seeing, the Fusion integration and deployment on OCI should hedge against the high AI failure rates that had recently been reported.

“Built-in AI is very different from DIY AI or bolted-on AI,” she said. “Built-in AI inherits so much from Fusion, like the security framework and guardrails.”

Wettemann said Oracle is investing in its AI capabilities around Studio to be competitive in the marketplace. “A key part of the adoption and deployment battle for AI is trust,” she said. “No one wants to unleash an AI agent on enterprise data without being sure it will act appropriately and deliver accurately. New observability and evaluation tools let users monitor latency, error rates and performance quality, while prompt management libraries help organize and manage versions of agents, which will help customers feel more comfortable with both initial deployment and ongoing performance.”

Image: SiliconANGLE/Microsoft Designer